
Multi-Tenant Architecture: 3 Models & 8 Best Practices
Learn how multi-tenant architecture works, compare database isolation models, and explore six components & eight best practices for shared infrastructure.
Consumption pricing ties charges to usage. See how it works, compare 6 models from pay-as-you-go to hybrid, and learn what engineers must build to run it.
A consumption-based pricing model charges customers for what they use, such as API calls, gigabytes, compute hours, or AI tokens. The more they use, the more they pay, and a slow month costs less.
It's how cloud providers have always billed, and AI products have made it the default for a new wave of software.
Customers like it because they pay for value they can see, and vendors like it because revenue grows as adoption grows.
The trade-off is that bills and revenue both get harder to predict, and when one AI agent can burn through a month’s budget overnight, someone has to control every unit in real time.
This guide covers how the model works, the six common types, the benefits and challenges, real examples, and how to tell whether it fits your product. It also covers what the system behind it needs, because that's where most consumption pricing runs into trouble.
A consumption-based pricing model charges customers based on their measured usage of a product, such as API calls, tokens, data stored, or compute time, so the bill rises and falls with how much they use. It's also called usage-based pricing or pay-as-you-go pricing.
An electricity bill is the everyday version. You pay for the kilowatt-hours you used this month, and if you use less next month, you pay less.
Flat subscriptions work the other way around, with the same price every month whether a customer uses the product heavily or barely touches it. Many software companies now blend the two, with a base subscription plus usage on top.
Consumption-based pricing works by choosing a unit of usage, measuring it per customer, applying a price to it, and billing for the total. In practice, that's four steps:
Ingestion collects raw events like API calls, token counts, storage writes, and agent actions. Metering turns them into the unit you price on, so storage activity becomes GB-months. Rating applies the pricing rules to produce charges. Getting the count right depends on usage metering architecture, from event capture to the pipeline that attributes each event to a customer.
Billing and enforcement both read from that pipeline, but in very different ways. Billing works on totals over time and can process yesterday's events this morning, while enforcement needs the current state during each request, which is a much harder requirement to meet.
The six main types of consumption-based pricing models are pay-as-you-go, tiered, volume, prepaid credits, overage, and hybrid.
| Model | How it works | What the system needs |
|---|---|---|
| 1. Pay-as-you-go | Charged per unit after usage | Accurate metering |
| 2. Tiered | Rate changes at usage thresholds | Tier tracking during rating |
| 3. Volume | Unit price drops as total usage grows | Aggregation across the period |
| 4. Prepaid credits | Usage draws down a prepaid balance | Real-time balance checks and a ledger |
| 5. Overage | Usage past a limit costs a different rate | Threshold detection and enforcement |
| 6. Hybrid | Subscription base plus usage | Plan access and usage checked together |
Each type has its own deep dive: pay-as-you-go, tiered pricing, credit pricing, overage pricing, and the hybrid pricing model.
Hybrid has become the most common choice because it balances predictability and flexibility. The subscription sets a baseline, and usage-based components handle the spikes, though it means plan terms and live usage have to be evaluated together on every request.
What you meter depends on how customers get value from the product, and each unit brings its own measurement problem.
The unit you choose shapes the system for years. Changing it later means migrating historical usage, updating pricing logic, and keeping existing customers' plans consistent. Snowflake’s CFO advises that the unit of consumption has to be simple for customers to understand.
Most of the best-known examples come from cloud infrastructure, data platforms, and AI, where the unit of value is easy to measure.
Here's what each one charges for, checked at each company's pricing page:
| Company | What it charges for |
|---|---|
| EC2 On-Demand charges for compute capacity by the hour or second, with a 60-second minimum, while Amazon S3 charges for storage per GB plus requests and data transfer. | |
| Compute is charged in credits billed per second, with a 60-second minimum each time a warehouse starts, and storage is billed at a flat monthly rate per terabyte. | |
| API usage is priced per 1 million tokens, with separate rates for input, cached input, and output on each model. | |
| API usage is also priced per million tokens, with input and output rates that differ by model and separate rates for prompt caching. | |
| Each paid plan includes a pool of model usage that refreshes weekly, and third-party models are billed separately on top of the plan. |
The pattern across all five is a clear unit, metered per customer, priced per unit, with an included amount or a commitment layered on top for predictability. Usage-based pricing examples walks through five common models in more detail.
Consumption-based pricing benefits both sides of the deal, which is a big reason it keeps spreading. The benefits just look different depending on who's paying.
The challenges of consumption-based pricing mostly come down to predictability, for both the vendor and the customer. Each one has a control that keeps it manageable.
Revenue moves with usage, so finance can't forecast it the way it forecasts a fixed subscription book. Minimum commitments, prepaid credits, and a subscription base give finance a floor to plan around, while usage data shows the trend above it.
A customer's spend can jump from one month to the next, which makes their finance team nervous. Budgets, spend caps, and per-team allocations let customers set limits they're comfortable with before they sign.
The fastest way to lose a consumption customer is an invoice nobody expected. Usage alerts as customers approach a threshold, and limits that enforce before usage runs, keep the bill in line with what the customer agreed to.
Deals often start smaller, because customers commit to less upfront, and reps paid on contract size lose out. Many companies move compensation toward consumption growth and net retention so reps are rewarded for usage that sticks.
In AI products, one heavy workflow or a misbehaving agent can burn through a budget in hours. It's contained with real-time limits and per-team budgets, the governance controls AI products add on top of metering.
Consumption-based pricing fits when the value customers get grows with how much they use, and when both sides can live with a bill that moves. These questions make the call easier:
If most answers are yes, consumption or hybrid pricing is worth testing. If value is flat regardless of usage, a subscription is often simpler for everyone.
Snowflake’s CFO takes a harder line and recommends committing to consumption, because a hybrid lets customers arbitrage the structure and makes revenue harder to predict.
AI products make consumption pricing harder because cost, usage patterns, and control requirements all change at once. Every token has a direct infrastructure cost, requests vary widely by model and input, and enterprise customers want budgets by team and workflow.
A single request to a reasoning model, or one multi-step agent run, can consume far more than expected. Without real-time enforcement, usage spikes before limits apply, and the cost lands as a surprise on someone's invoice.
That's why AI products add a governance layer that sits in the request path:
Enterprise AI customers need these controls at several levels at once, with per-agent caps, per-team allocations, and department spend limits all enforcing independently. A flat per-customer balance was never built for that.
Consumption-based systems break when usage, state, and enforcement fall out of sync under real load. A simple setup works at small scale, and the gaps show up as concurrency grows and plans evolve.
| Failure mode | What happens |
|---|---|
| Concurrent usage conflicts | Several requests update one balance at once |
| Stale state | Cached usage or entitlements fall behind |
| Delayed enforcement | Usage is recorded after the request completes |
| Wrong attribution | Agent steps land on the wrong customer or feature |
| Mid-cycle plan changes | New limits don't reach active sessions |
| No audit trail | Balances update in place with no event history |
The audit trail is the failure that shows up last and costs the most to fix. A mutable balance column tells you where a balance ended up, while a ledger records every grant and deduction along the way, which is what dispute resolution and revenue recognition both depend on.
At this stage, these problems usually trace back to design: metering, enforcement, and state were never built to stay in sync.
Consumption pricing infrastructure should grow with the company. Early stages need basic metering, growth stages add tiers and enforcement, and late stages add governance and credit control.
| Company stage | Pricing focus | What engineering needs to build |
|---|---|---|
| Early stage | One measurable usage unit | Event capture, usage attribution, reliable counting |
| Growth stage | Pricing tiers and hybrid models | Entitlements enforcement, tier thresholds, configurable pricing |
| Late stage | Governance and credit control | Credit ledgers, usage limits, real-time enforcement |
Pick one usage metric that maps to product value and build metering around it. A storage product counts gigabytes, an API counts requests, and an AI product counts tokens.
Early-stage metering needs three pieces:
Tiers, overages, and credit ledgers can wait until that first unit counts accurately.
Growth-stage companies juggle multiple tiers, a mix of flat and variable pricing, and customers at different usage levels. The in-house system built early starts to strain, because every pricing change touches billing logic, product behavior, and access rules at the same time.
Entitlements enforcement becomes its own layer, and packaging changes such as new limits or tiers move through the product catalog with no code deployment.
Late-stage companies hit governance problems. A power user consumes a disproportionate share of resources, a usage spike produces a disputed invoice, and an enterprise customer asks to split credits across departments.
The entitlements layer has to enforce limits and credit policies before the usage event completes, because an invoice arrives too late to stop anything.
Billing and enforcement are separate jobs, and treating them as one system is where many consumption-based architectures go wrong.
| | Billing | Enforcement |
|---|---|---|
| When it runs | After consumption | Before consumption |
| What it reads | Usage totals over time | Current usage state |
| Tolerates delay | Yes | No |
| Main concern | Invoicing and revenue recognition | Access and limits |
Billing tools are precise and good at recording consumption and generating invoices, and our consumption-based billing guide covers those mechanics. What they don't do is decide inside the product, at request time, whether the next request should run.
That decision belongs in a layer upstream of billing. Stripe handles payments, and the enforcement layer handles what's allowed. For the step-by-step rollout, our usage-based pricing guide walks through implementation.
When consumption pricing does fit, the pricing model is usually the simple part, and keeping usage under control at request time is where the work goes.
Stigg is a usage runtime that sits between your product and your billing system, resolving entitlements, credits, and spend limits in the request path before compute runs. It works alongside your billing, with native integrations for Stripe, Zuora, and Airwallex.
Miro launched AI credits on Stigg in under six weeks, with free and paid users getting different monthly token allotments that are tracked, enforced, and reset automatically.
With Stigg, engineering teams can:
If your pricing already depends on usage and enforcement still happens after the fact, that's the gap worth closing first. The Stigg docs show how it fits next to your billing, or you can talk to a human about your stack.
The four pricing models most software companies choose between are flat-rate, seat-based, tiered, and consumption-based pricing, and many combine two of them.
| Model | How customers pay |
|---|---|
| Flat-rate | One price for the whole product |
| Seat-based | A price per user |
| Tiered | Set packages with different features or limits |
| Consumption-based | A price per unit of usage |
The main difference between consumption pricing and subscription pricing is how the bill is calculated. Subscription pricing charges a fixed fee at any usage level, while consumption pricing charges in proportion to usage and needs accurate metering to do it.
Yes, the two terms are used interchangeably. Both charge customers based on measured usage, and "consumption-based" tends to show up more in cloud and enterprise contracts, while "usage-based" is more common in SaaS.
The main difference between consumption pricing and usage-based billing is the layer each term describes. Consumption pricing is the commercial model, and usage-based billing is the system that counts usage and generates invoices from that data.
Companies capture product events such as API calls, tokens, or compute time and convert them into billable units in a metering pipeline. Every event needs a unique identifier, so retries and duplicate processing don't inflate the count.
The system either blocks further usage or lets it continue and records an overage, depending on the plan. That decision has to be made during the request, against the current balance, so the limit holds when it matters.
Teams set limits, allocate budgets by team or agent, and enforce those controls in the request path. Alerts help with visibility, and the limit that runs before the request is what stops the spend.
Yes, for any product where usage carries real cost or customers expect hard limits. A system that meters accurately but enforces later will record every event correctly while still letting usage run past the limit, which is how overages and billing disputes start.